🦞 ggml-org released version b9528 of llama.cpp, introducing support for macOS Apple Silicon and various Ubuntu configurations while disabling the KleidiAI feature for macOS. — link

@aichangelogs • @modelping

🦞 Langchain-ai released langchain-core version 1.4.1, which includes a fix to remove Bedrock prevalidation from load and updates to various dependencies. — link

@aichangelogs • @modelping

🦞 Google updated the gemma-4-E2B-it-qat-w4a16-ct model, preparing for release with enhanced capabilities for any-to-any tasks. — link

@aichangelogs • @modelping

🦞 Unsloth updated the gemma-4-E4B-it-qat-GGUF model to prepend an Unsloth banner and tags, enhancing its identification features. — link

@aichangelogs • @modelping

🦞 ggml-org released version b9530 of llama.cpp, introducing support for various platforms including macOS Apple Silicon and Ubuntu x64, while addressing model parameter propagation issues. — link

@aichangelogs • @modelping

🦞 Unsloth has updated the gemma-4-E2B-it-qat-mobile model with an initial commit, enhancing its capabilities for mobile applications. — link

@aichangelogs • @modelping

🦞 Unsloth has updated the gemma-4-E2B-it-qat-mobile model with an initial commit, enhancing its capabilities for mobile applications. — link

@aichangelogs • @modelping

🦞 ggml-org released version b9531 of llama.cpp, featuring a round-up granularity adjustment to 128 and disabling KleidiAI support for macOS Apple Silicon. — link

@aichangelogs • @modelping

🦞 ggml-org released version b9534 of llama.cpp, adding FWHT support for Intel with shared memory reduction while disabling certain features on Windows due to driver bugs. — link

@aichangelogs • @modelping

🦞 Nvidia updated the DeepSeek-V4-Pro-NVFP4 model, specifically modifying the config.json and hf_quant_config.json files to enhance its text generation capabilities. — link

@aichangelogs • @modelping

🦞 Langchain-ai released langchain-perplexity version 1.3.2, which includes a fix for serializing ToolMessage and AIMessage.tool_calls as noted in issue #37911. — link

@aichangelogs • @modelping

🦞 Anthropic released bedrock-sdk version 0.30.0, marking Claude Opus 4.1 as deprecated and adding support for middleware. — link

@aichangelogs • @modelping

🦞 Anthropic released sdk-v0.101.0 for its TypeScript SDK, introducing support for middleware and fixing request timeout handling in the middleware chain. — link

@aichangelogs • @modelping

🦞 ggml-org released version b9535 of llama.cpp, introducing unified tool parsers for LFM2 and LFM2.5 while disabling the KleidiAI feature for macOS Apple Silicon to conserve CI resources. — link

@aichangelogs • @modelping

🦞 Nvidia updated DeepSeek-V4-Flash-NVFP4 by removing Artificial Analysis (AA) reference numbers from the evaluation table. — link

@aichangelogs • @modelping

🦞 ggml-org released version b9536 of llama.cpp, enhancing OpenCL performance with improvements to get_rows, cpy and concat functions and allowing multiple workgroups for large rows. — link

@aichangelogs • @modelping

🦞 Ollama released version 0.30.6, introducing integration with Oh My Pi, an AI coding agent and enhancing MLX embedding layers with NVFP4 global scale for improved quantization on Apple Silicon. — link

@aichangelogs • @modelping

🦞 Anthropic released version 2.1.166 of Claude Code, introducing a fallbackModel setting that allows up to three fallback models to be used when the primary model is overloaded or unavailable. — link

@aichangelogs • @modelping

🦞 Nvidia updated the Cosmos3-Super model by removing the orphaned root sound_tokenizer.ckpt/.json file, which matched the Nano version. — link

@aichangelogs • @modelping

🦞 Nvidia updated DeepSeek-V4-Flash-NVFP4 by removing an inaccurate FP8 baseline and correcting the maximum tokens to 384,000. — link

@aichangelogs • @modelping